Dony Kushardono

Comparison of multi-temporal image classification methods

One of the promising methods which can be thought to increase classification accuracies in remote sensing is the use of multi-temporal images. The authors propose multi-temporal image classification methods using backpropagation networks and fuzzy neural networks as classifiers and two kinds of classification models based on co-occurrence matrix as spatial information source. They are compared with conventional methods such as the likelihood addition method, the likelihood majority method and the Dempster-Shafer rule method.

Spatial land cover classification with the aid of neural network

A land cover classification method using a neural network was applied for the purpose of utilizing spatial information, which is expressed as a two-dimensional array of a co-occurrence matrix. The adopted neural network has three layers feed forward network architecture with back-propagation learning algorithm. In this study, the three kinds of neural network classification models were proposed. The first and the second model classifies each band image at the first stage, then performs final decision based on the first stage result. At the decision stage, arithmetic decision algorithm and second neural network are used by the first and the second model, respectively. The third model is a single stage classifier that enters all band information into the neural network for learning and classification at the same time. In order to evaluate proposed models, land cover classification using the proposed models and conventional pixel wise maximum likelihood method was conducted with Landsat TM and SPOT HRV data. As a result, the third model showed best performance, with accuracies about 4% to 6% higher than those of the classification result of the first and second model, and it showed about 17% to 27% higher than that of the maximum likelihood classification result. Finally, we examine the best performance of the neural network classification model for multitemporal remote sensing data classification, which was successful.

Performance of LAPAN-A2 satellite data to classify land cover/land use in Semarang, Central Java

LAPAN-A2 is a 2nd generation of microsatellite developed by the Indonesian National Institute of Aeronautics and Space (LAPAN), Indonesia. This satellite has been launched on September 28, 2015, with several primary missions are: Earth observation using an RGB camera, maritime traffic monitoring using Automatic Identification System (AIS), and supporting disaster mitigation through amateur radio automatic packet reporting system (APRS) and a voice repeater. The installed camera on the satellite can provide imagery with 3.5 m spatial resolution and able to view a swath 7 km wide. This study is attempt to classify and evaluate the accuracy of land cover and land use (LCLU) map classified from the LAPAN-A2 satellite data. As a reference we used the Pleiades-1A Orthorectified imagery data with 0.5 m spatial resolution. Through visual interpretation method, we classified the LCLU into five classess as follows: water body, building, vegetation, road, and bare land. The confusion matrix was applied to evaluate the classification results for three selected area of interest (AOI). We obtained the overall accuracy, producers accuracy, and users accuracy are about 61.77%, 61.02%, and 83.78% respectively.